Deep Learning and Blockchain Improve Secure Vehicle Routing

Researchers have proposed a new framework called Deep Learning-based Adaptive Routing with Blockchain-enabled Trust Management, or DL-ARBTM, to address the security and efficiency challenges in Vehicular Ad Hoc Networks. Published in Discover Internet of Things, the study introduces a two-layered mechanism that integrates Deep Reinforcement Learning for routing decisions with a blockchain-based system for managing node trust. The primary goal is to create a communication network for Intelligent Transportation Systems that can withstand malicious attacks while maintaining high data delivery rates in highly dynamic traffic environments.

Vehicular Ad Hoc Networks are specialized mobile networks that enable real-time communication between vehicles and roadside infrastructure. These networks are essential for applications such as collision avoidance, cooperative driving, and dynamic route planning. However, the high mobility of vehicles creates a constantly changing network topology, which makes traditional routing protocols like AODV and GPSR less effective. These conventional methods often struggle with latency and packet loss when the network structure shifts rapidly due to vehicle movement.

The DL-ARBTM framework addresses these limitations by modeling the routing process as a Markov Decision Process. In this model, each vehicle acts as an autonomous agent that monitors link quality, node mobility, and network density. Using an Actor-Critic Deep Reinforcement Learning agent, the system dynamically selects the best next-hop node for data forwarding. This approach allows the network to adapt in real-time to changing conditions, such as rapid changes in link availability or channel interference, which static protocols cannot handle effectively.

Security is a critical concern in these decentralized networks, where malicious nodes may drop packets, spread misinformation, or perform Sybil attacks. To counter these threats, the framework incorporates a blockchain layer that maintains an immutable, distributed registry of node trust scores. This system uses Practical Byzantine Fault Tolerance consensus to ensure that trust updates are validated by a majority of honest validators. By recording node behavior on a tamper-proof ledger, the network can identify and isolate malicious actors without relying on a centralized authority that could become a single point of failure.

The integration of these two layers creates a closed-loop system where trust directly influences routing. The Deep Reinforcement Learning agent uses the blockchain-validated trust score as part of its state vector and reward function. This ensures that the agent only forwards data through nodes that have a verified trust score above a specific threshold. Consequently, the routing process becomes not only adaptive to physical network conditions but also resilient to adversarial behavior, as untrusted nodes are effectively excluded from the data path.

The study evaluated the performance of DL-ARBTM through extensive simulations using the SUMO mobility simulator and a custom Gymnasium environment. The simulation environment was a 5 km by 5 km urban grid with vehicle densities ranging from 50 to 300 vehicles per square kilometer. Vehicles moved at speeds between 30 and 80 km/h, and the communication was modeled using the IEEE 802.11p standard. The researchers varied the proportion of malicious nodes from 0% to 30% to test the framework’s resilience under different threat levels.

In terms of packet delivery ratio, the DL-ARBTM framework achieved an average of 95.35%, significantly outperforming the baseline protocols. The blockchain-only approach achieved a packet delivery ratio of 88.1%, while the Q-learning baseline reached 86.3%. Traditional protocols performed the worst, with AODV at 82.5% and GPSR at 78.9%. These results suggest that the combination of adaptive learning and decentralized trust management is crucial for maintaining reliability in high-mobility vehicular environments where static routing methods fail.

Latency performance was also a key metric in the evaluation. DL-ARBTM demonstrated the lowest end-to-end latency at 102.42 milliseconds, which is critical for safety-critical applications that require immediate data exchange. In comparison, the blockchain-only approach had a latency of 115.9 milliseconds, and the Q-learning baseline recorded 108.7 milliseconds. The traditional protocols AODV and GPSR showed higher latencies of 120.5 and 135.2 milliseconds, respectively. The lower latency in the proposed framework is attributed to the DRL agent’s ability to quickly select the most efficient and trusted paths.

Throughput measurements further highlighted the efficiency of the proposed system. DL-ARBTM achieved a maximum throughput of 952 kbps, surpassing the blockchain-only variant at 890 kbps and the Q-learning baseline at 860 kbps. The traditional protocols AODV and GPSR recorded the lowest throughputs at 820 kbps and 780 kbps. This superior data delivery rate indicates that the framework can handle high volumes of traffic while maintaining security, a vital requirement for future intelligent transportation systems that rely on constant data exchange.

The study also examined the impact of malicious nodes on trust accuracy. As the proportion of malicious nodes increased to 30%, the average trust score in the DL-ARBTM model decreased from 0.98 to 0.82. In contrast, the blockchain-only approach saw its trust score drop from 0.95 to 0.73, and the Q-learning baseline fell from 0.90 to 0.65. This demonstrates that the adaptive nature of the DRL component helps the network maintain a higher level of trust integrity even under significant adversarial pressure, where other methods struggle to distinguish between honest and malicious behavior.

Routing overhead was measured to assess the additional control traffic required by the framework. DL-ARBTM had a maximum overhead of 27.3%, which is higher than the blockchain-only approach at 21.5% and the Q-learning baseline at 20.5%. The traditional protocols AODV and GPSR had lower overheads at 18.4% and 19.2%. While the increased overhead is a trade-off for the enhanced security and reliability, the researchers argue that this is a reasonable cost for safety-critical applications where data integrity is paramount. The overhead is primarily due to the blockchain transactions and the periodic updates required by the reinforcement learning agent.

An ablation study was conducted to isolate the contributions of the DRL and blockchain components. The results showed that the DRL-only variant had low latency but poor trust accuracy under attack, while the blockchain-only variant was secure but suffered from higher latency and lower throughput due to the lack of adaptive routing. Only the full hybrid DL-ARBTM achieved the best performance across all metrics simultaneously. This confirms that the two components are complementary and that neither alone is sufficient to provide a robust solution for secure vehicular communication.

The framework also addresses scalability concerns by decoupling the routing decisions from the global network size. The DRL agent makes local decisions based on immediate neighbors, keeping the computational cost linear with the number of neighbors rather than the total number of vehicles. The blockchain consensus is handled by a fixed set of Roadside Unit validators, ensuring that the validation cost does not grow with the vehicle population. This design allows the system to remain efficient and responsive even as the number of connected vehicles increases in large-scale deployments.

Despite its advantages, the study acknowledges certain limitations. The trust evaluation is behavioral and reactive, meaning that a node that behaves honestly before defecting may still succeed for one trust-update interval. Additionally, the security guarantees hold only within the Byzantine bound of the consensus mechanism. The researchers note that an empirical sweep beyond 30% malicious nodes is left for future work, as the current model assumes that the fraction of compromised validators remains below the threshold required to break the consensus. This framework provides a promising path toward more secure and efficient communication in the evolving landscape of intelligent transportation systems.

Subject of Research: Secure adaptive routing in Vehicular Ad Hoc Networks using deep learning and blockchain

Article Title: A deep learning and blockchain enabled adaptive routing framework for secure Vehicular Ad Hoc Networks

Article References: Oli, J. M., Neelima, N., Vasquez, J. C., & Bahadur, J. (2026). A deep learning and blockchain enabled adaptive routing framework for secure Vehicular Ad Hoc Networks. Discover Internet of Things, 6(1), Article 185. https://doi.org/10.1007/s43926-026-00516-2

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00516-2

Keywords: Vehicular Ad Hoc Networks, Deep Reinforcement Learning, Blockchain, Intelligent Transportation Systems, Adaptive Routing, Network Security, Trust Management, PBFT Consensus, deep, learning, enabled, adaptive